Skip to content
Help

AI grading

AI can draft a grade, check yours, or stay out of it entirely. Whichever you choose, a person always makes the final decision — that is not a setting.

The four modes

Section titled “The four modes”

Set per assignment on the Schedule stage, starting from your organization’s default.

Mode What happens
People only No AI. The grader scores every hand-in.
AI drafts, you approve The assistant drafts a grade and feedback first; you accept, edit or reject each one — or accept all.
AI drafts, you verify each The assistant drafts; every grade must be checked and confirmed by a person before release.
You grade, AI checks You grade first; the assistant gives an independent second opinion and flags large disagreements.

The difference between the second and third is whether bulk acceptance is available. The fourth inverts the order — your judgement comes first, and AI is a check on it rather than a starting point.

What the AI actually reads

Section titled “What the AI actually reads”

It reads the hand-in:

  • PDFs and images are read directly.
  • Word, PowerPoint and text files are read as extracted text, with personal details stripped before they’re sent.
  • Audio and video can’t be read. There’s no transcription, so the assistant says it couldn’t read the work rather than guessing at it.

Where it could only read part of a hand-in, the draft says so: “The assistant could not read part of the hand-in — check the file yourself before deciding.” Treat that as a prompt to open the file.

A learner’s name and email are never sent.

AI drafts, you approve

Section titled “AI drafts, you approve”

A draft arrives per hand-in, and the roster row reads AI graded — awaiting your decision.

The drawer’s draft card shows the score per criterion, why the assistant scored it that way, the evidence it’s relying on — a quotation where the criterion is about content, an observation where it’s about the work as a whole — and a confidence figure. Where a criterion falls short it says what would score higher.

Then:

  • Use these scores copies them into the form so you can adjust before grading.
  • Accept AI draft takes them as they are.
  • Accept all AI grades does that for every unreviewed draft at once. “You can still open any of them and change the grade.”

Nothing reaches the learner until you release it, and release is blocked while an AI draft is still unapproved.

You grade, AI checks

Section titled “You grade, AI checks”

You grade normally. The assistant then scores independently and compares:

  • Agrees with you (4% apart) — nothing to do.
  • Diverges by 22% — moderation flagged — the grade is held for moderation.

The panel shows both scores side by side per criterion. Keep mine stands by your grade; Adopt AI scores copies theirs into the form for you to review before submitting. You can ask again with notes saying what the second opinion should focus on, a limited number of times.

Your decision is final. Once you’ve resolved a check, later checks never re-raise the hold. This matters because the model is not deterministic — the same hand-in can score differently on consecutive runs, and a system that kept re-flagging would never let a grade out.

What the learner sees

Section titled “What the learner sees”

On a released grade where AI helped:

An AI assistant helped draft this grade. A person made the final decision.

This is a legal transparency requirement, not a preference, and it can’t be switched off.

Learners can opt out

Section titled “Learners can opt out”

AI grading is an optional consent in a learner’s privacy settings. Someone who declines is graded by a person, and the roster row reads Human review (learner opted out of AI).

The assignment itself tells them before they hand in:

When AI helps grade this assignment, a person always decides the grade. Your name and email are never sent; you can opt out in Privacy settings.

Setting the organization default

Section titled “Setting the organization default”

Someone with permission to manage AI settings picks the default new assignments start from — “Teachers can change it per assignment.”

Some organizations, particularly those working with children, require a second approver before AI grading can be switched on at all: the request records the lawful basis, a different person approves it, and enabling it is written to the audit trail.

Reference

Section titled “Reference”
Thing Behaviour
Final decision Always a person, in all four modes
Release Blocked while an AI draft is unapproved
Readable PDF, images, Word, PowerPoint, text
Not readable Audio, video — it says so rather than guessing
Sent to the model The work. Never the learner’s name or email
Learner opt-out An optional consent; graded by a person instead
Disclosure Shown on every released AI-assisted grade. Not switchable
Scoring Against the rubric’s maximum, pinned to the rubric version

Troubleshooting

Section titled “Troubleshooting”

The AI grading panel is missing. AI isn’t switched on for your organization, or the settings couldn’t load — in both cases grading falls back to people only.

Release is greyed out. An AI draft is still awaiting your decision, or a moderation hold is open. Resolve it and release.

The same hand-in scored differently twice. The model isn’t deterministic. That’s why a person decides, and why your resolved decision is never re-flagged.

The score looks wrong against the points. An AI record scores against the rubric’s maximum. If the assignment’s points and the rubric’s maximum differ, read the rubric total.

It said it couldn’t read the file. Audio and video can’t be read at all. For anything else, check the file opens — a corrupt or image-only PDF may defeat it.

A learner’s row says they opted out. They withdrew AI-grading consent. Grade it yourself; nothing is wrong.

Section titled “Related”